Qwen3.8-Max: Open Weights Hit Data-Center Speed on NVIDIA, AMD

Qwen3.8-Max Hits Production Speed on GB300 NVL72

The largest open-weight release from Alibaba is no longer a research artifact. It is running at data-center velocity.

NVIDIA’s technical team has published Day-0 serving data for Qwen3.8-2.4T-A95B, the 2.4-trillion-parameter model also known as Qwen3.8-Max, on the GB300 NVL72 rack-scale platform.

In FP8 precision, the system exceeds 4,000 tokens per second per GPU and more than 350 tokens per second per user without additional tuning.

The performance lands on a model that keeps only 95 billion parameters active per token, making the full 2.4T-parameter architecture far more practical for demanding production workloads than a dense model of equivalent scale.

That combination, detailed in the NVIDIA developer blog, places Qwen3.8-Max at the center of the open-weight race: near-frontier capability, million-token context, and a launch-day serving story on multiple hardware ecosystems.

Hybrid Attention and Fine-Grained MoE Redesign Long-Context Serving

Qwen3.8-2.4T-A95B is tuned for agentic workloads that accumulate large amounts of context: system instructions, retrieved documents, tool outputs, logs, and multi-step reasoning traces.

In those workflows, full-attention layers and traditional KV cache sizes often become the limiting constraint.

The model’s architecture alternates between full attention, where every token attends to every other token, and linear attention, where the expanding cache is replaced with a bounded recurrent state.

This hybrid approach keeps compute and memory bounded as context scales toward one million tokens.

The model also uses a fine-grained mixture of experts rather than a small set of large expert blocks.

A learned router activates only the experts required per token, so serving cost follows the 95 billion active parameters rather than the full 2.4 trillion.

Built-in reasoning controls, labeled low, high, and xhigh, let developers trade compute for reasoning depth per request.

On the GB300 NVL72, 72 Blackwell Ultra GPUs sit inside one NVLink domain with 130 terabytes per second of all-to-all fabric.

That design removes the cross-node communication stalls that typically appear when expert traffic has to traverse conventional networks.

For serving, developers can choose SGLang, vLLM, or NVIDIA Dynamo, while the model-free NVIDIA NIM container offers a packaged path to production.

Weights are available through Hugging Face and ModelScope, and post-training can start through NeMo AutoModel without checkpoint conversion.

AMD’s Instinct Day-0 Counter and the Open-Weights Silicon Race

AMD’s Day-0 support announcement for the same Qwen 3.8 family arrived the same day, covering Instinct MI300X, MI325X, and MI355X GPUs through the ROCm software stack.

That simultaneous support on both major AI silicon ecosystems is a clear escalation in the open-weights race.

The support note describes the model as a sparse Mixture-of-Experts design with 512 experts and a 92-layer network, with reasoning effort and preserved thinking history available through reasoning_effort and preserve_thinking.

Deployment paths include SGLang, vLLM, and ATOM across ROCm, with checkpoints on Hugging Face and ModelScope.

On metrics, a vLLM FP8 run across two MI355X nodes with 16 total GPUs delivered GSM8K exact match of 0.9591 ± 0.0055.

A single-node MI355X MXFP4 vLLM configuration hit a strict-match score of 0.9568, while SGLang with MXFP4 reached roughly 1,060 output tokens per second.

SGLang FP8 on two MI300X nodes managed 112.7 output tokens per second, a lower figure that reflects the older-generation accelerator rather than a flaw in the model.

ATOM MXFP4 on one MI355X reported a strict exact match of 94.92 percent and flexible exact match of 95.00 percent.

These are vendor-published benchmark results, and they have not been independently validated under identical serving conditions across hardware families.

Community rankings added another caveat. Some early discussions positioned Qwen3.8-Max as the top overall model, but a widely shared screenshot appeared to show Claude Opus 5 ahead by 0.8 points on the Artificial Analysis Agentic Index.

That gap places Qwen3.8-Max in the top tier rather than an undisputed number one, which tempers some of the more exuberant launch-day claims.

What matters for infrastructure teams is not a single leaderboard slot. It is that a 2.4T-parameter open model can now be tuned, quantized, and served across different silicon choices without waiting for an upstream ecosystem to catch up.

What Open Frontier Throughput Unlocks for AI Teams

Day-0 support on both major silicon ecosystems turns Qwen3.8-Max from a model drop into an infrastructure decision point. For AI teams building agentic pipelines that need to scale, Andres SEO Expert’s programmatic SEO and AI automation service is how technical authority becomes search-visible — contact us here.